Crosstalk estimator
How much of the aggressor lands on the victim, at both ends — and whether the 3W rule is doing what you think on this stack-up.
Worst-case coupled noise
—mV
—
The physics
Coupling is a ratio of geometries
k ≈ 1 / (1 + (D / H)²) D = centre-to-centre, H = height above planeThis is the whole story in one line. Coupling depends on spacing relative to height above the reference plane, not on spacing alone. A pair of traces 0.4 mm apart over a plane 0.1 mm below them barely couple; the same pair over a plane 0.5 mm below couple badly. Bringing the plane closer is often easier than spreading the traces, and it fixes impedance control at the same time.
Near end and far end behave differently
V_NEXT = k_b · V_swing (saturates with length)
V_FEXT = k_f · V_swing · (L / t_r) (grows with length)Backward crosstalk travels toward the aggressor's driver and stops growing once the parallel run exceeds the saturation length — half the physical length of the rising edge. Forward crosstalk accumulates along the whole run and gets worse with faster edges. Long runs hurt the far end; short fast ones hurt the near end.
Why stripline is quieter
Far-end crosstalk is proportional to the mismatch between the inductive and capacitive coupling terms. In a homogeneous dielectric they cancel and FEXT nearly vanishes. Microstrip has air on one side and laminate on the other, so it does not cancel. If far-end noise is your problem, moving to an inner layer is a real fix, not a marginal one.
Edge rate, not clock rate
A 1 MHz signal with a 200 ps edge couples like a fast signal, because coupling
responds to dV/dt. Modern logic families have edges far faster than
their rated frequency requires. If a slow control line is causing trouble, a series
resistor at the driver to slow the edge deliberately is often the cheapest fix
available.
Where this stops being accurate
- These are closed-form approximations. A 2D field solver is what you use when the answer is close to the margin.
- Only two conductors are modelled. Multiple aggressors are scaled linearly here, which is pessimistic for random data and optimistic for a bus switching in unison.
- Assumes a continuous reference plane. A split or a gap under the run raises coupling dramatically and is a far more common failure than tight spacing.
- Termination is assumed ideal. Reflections on either line change both numbers.
- No via or connector coupling, which frequently exceeds trace coupling in the same design.
Design guidance, not a manufacturing instruction. Verify every result against your fabricator's stack-up and the applicable standard before release. Closed-form models are approximations; the fab's field solver and process window are the authority.
Questions
What people ask about this
Is the 3W rule good enough?
It is a decent default and a poor law. 3W — centre-to-centre spacing three times the trace width — typically gets coupling under a few percent for ordinary geometry. But it says nothing about height above the plane, which matters as much as spacing, and nothing about run length or edge rate. Two traces at 3W over a distant plane couple far worse than two at 2W hugging a close one. Check the number, not the rule.
Why is the near-end noise at the driver, not the victim?
Backward crosstalk propagates toward the source end of the victim line, so it appears at the end nearest the aggressor's driver. It also saturates: once the coupled region is longer than half the rise-time length, near-end noise stops growing with length. That is why a long parallel run does not keep getting worse at the near end — but the far end does.
Why does far-end crosstalk vanish in stripline?
Far-end crosstalk is proportional to the difference between inductive and capacitive coupling. In a homogeneous dielectric — stripline, with the same material above and below — those two terms cancel almost exactly and far-end coupling approaches zero. Microstrip has air above and laminate below, so they do not cancel, and far-end noise is often the dominant problem. If far-end crosstalk is hurting you, moving the layer inward is a genuine fix.
What coupling level is acceptable?
Depends entirely on the victim's noise margin. A CMOS input with 400 mV of margin tolerates a lot; a differential receiver with 100 mV does not; an ADC reference or a reset line tolerates almost nothing. A useful working target is keeping total coupled noise under 5% of the signal swing for logic, and under 1% for anything analog or asynchronous. Reset, interrupt and clock enable lines deserve the tightest budget because a glitch there is not recoverable.
Intermittent faults that move when you touch the board?
Crosstalk failures are rarely reproducible and always expensive to find late. We review the stack-up and the routing before the fab order goes out.